Anthropic's Safety Roadmap: When AI Gets Too Smart to Ignore

Article written and saved to src/content/articles/anthropic-responsible-scaling-policy-2026.md.

Here’s what the piece covers (~1,450 words):

• Opening — frames the RSP not as PR but as a capability roadmap hiding in plain sight
• ASL levels — concrete breakdown of ASL-1 through ASL-4+, including the Version 3.1 clarification that the AI R&D threshold means “compressing two years of 2018-2024 AI progress into a single year”
• Version 3.0 deep dive — Frontier Safety Roadmaps and Risk Reports as structural additions, not cosmetic ones
• Sabotage assessment — why the Claude Opus 4.6 findings matter, and why the admission of measurement difficulty is the more important sentence
• How to Use It — four practical applications: enterprise procurement, reading new model releases, accountability tracking, capability trajectory inference
• Competitor comparison — OpenAI Preparedness Framework, Google DeepMind Frontier Safety Framework, Meta’s looser approach, with honest differentiators
• Honest take — split between what’s genuinely impressive (threshold specificity, roadmap accountability), what’s concerning (self-certification, May 2025 scope narrowing), what’s overhyped, and what’s undersold (org structure changes)
• Conclusion — reframes the RSP as a capability roadmap, not just a safety document